better templates

api/src/stampy_chat/chat.py

@@ -223,34 +223,8 @@ def create_outputs(self, llm_result) -> List[Dict[str, Any]]:
         return [dict(self.inputs, **r) for r in result]
 
 
-def make_history_summary(settings):
-    model = get_model(
-        streaming=False,
-        max_tokens=settings.maxHistorySummaryTokens,
-        model=settings.completions
-    )
-    summary_prompt = PrefixedPrompt(
-        input_variables=['history'],
-        messages_field='history',
-        prompt=settings.history_summary_prompt,
-        transformer=ChatMessage,
-    )
-    return LLMInputsChain(
-        llm=model,
-        verbose=False,
-        output_key='history_summary',
-        prompt=ModeratedChatPrompt.from_messages([
-            summary_prompt,
-            ChatPromptTemplate.from_messages([
-                HumanMessagePromptTemplate.from_template(template='Q: {query}', role='user'),
-            ]),
-            HumanMessage(content="Reply in one sentence only"),
-        ]),
-    )
-
-
 def make_prompt(settings, chat_model, callbacks):
-    """Create a proper prompt object will all the nessesery steps."""
+    """Create a proper prompt object with all the necessary steps."""
     # 1. Create the context prompt from items fetched from pinecone
     context_template = "\n\n[{{reference}}] {{title}} {{authors | join(', ')}} - {{date_published}} {{text}}\n\n"
     context_prompt = MessageBufferPromptTemplate(
@@ -269,8 +243,7 @@ def make_prompt(settings, chat_model, callbacks):
         [
             HumanMessage(content=settings.question_prompt),
             HumanMessagePromptTemplate.from_template(
-                template='{history_summary}{delimiter}{query}',
-                partial_variables={"delimiter": lambda **kwargs: ": " if kwargs.get("history_summary") else ""}
+                template=f'{settings.question_marker}: {{query}}',
             ),
         ]
     )
@@ -344,32 +317,20 @@ def run_query(session_id: str, query: str, history: List[Dict], settings: Settin
         model=settings.completions
     )
 
-    history_summary_chain = make_history_summary(settings)
-
-    if history:
-        history_summary_result = history_summary_chain.invoke({"query": query, 'history': history})
-        history_summary = history_summary_result.get('history_summary', '')
-    else:
-        history_summary = ''
-
-    delimiter = ": " if history_summary else ""
-
     llm_chain = LLMChain(
         llm=chat_model,
         verbose=False,
         prompt=make_prompt(settings, chat_model, callbacks),
         memory=make_memory(settings, history, callbacks)
     )
 
-    chain = history_summary_chain | llm_chain
+    chain = llm_chain
     if followups:
         chain = chain | StampyChain(callbacks=callbacks)
 
     chain_input = {
         "query": query,
         'history': history,
-        'history_summary': history_summary,
-        'delimiter': delimiter,
     }
 
     result = chain.invoke(chain_input)

api/src/stampy_chat/followups.py

@@ -76,7 +76,7 @@ class Config:
 
     @property
     def input_keys(self) -> List[str]:
-        return ['query', 'text', 'history_summary']
+        return ['query', 'text']
 
     @property
     def output_keys(self) -> List[str]:

api/src/stampy_chat/settings.py

@@ -9,28 +9,33 @@
 
 SOURCE_PROMPT = (
     "You are a helpful assistant knowledgeable about AI Alignment and Safety. "
-    "Please give a clear and coherent answer to the user's questions. (written after \"Q:\") "
+    "Please give a clear and coherent answer to the user's questions. (written after \"Question:\") "
     "using the following sources. Each source is labeled with a number. Feel free to "
     "use the sources in any order, and try to reference up to 8 sources in your answers.\n\n"
+    "# Sources\n"
 )
 HISTORY_PROMPT = (
     "\n\n"
-    "Before the question (\"Q: \"), there will be a history of previous questions and answers. "
+    "# History:\n\n"
+    "Before the question (\"Question:\"), there will be a history of previous questions and answers. "
     "These sources only apply to the last question. Any sources used in previous answers "
     "are invalid."
 )
 HISTORY_SUMMARIZE_PROMPT = (
     "You are a helpful assistant knowledgeable about AI Alignment and Safety. "
-    "Please summarize the following chat history (written after \"H:\") in one "
-    "sentence so as to put the current questions (written after \"Q:\") in context. "
+    "Please summarize the following chat history (written after \"History:\") in one "
+    "sentence so as to put the current questions (written after \"Question:\") in context. "
     "Please keep things as terse as possible."
-    "\nH:"
+    "\nHistory:"
 )
 
 QUESTION_PROMPT = (
+    "# Question context:\n\n"
     "In your answer, please cite any claims you make back to each source "
     "using the format: [1], [2], etc. If you use multiple sources to make a claim "
-    "cite all of them. For example: \"AGI is concerning [1, 3, 8].\"\n\n"
+    "cite all of them. For example: \"AGI is concerning [1, 3, 8].\"\n"
+    "Don't explicitly mention the sources unless it impacts the flow of your answer - just cite "
+    "them. Don't repeat the question in your answer. \n\n"
 )
 PROMPT_MODES = {
     'default': "",
@@ -57,6 +62,7 @@
     'history_summary': HISTORY_SUMMARIZE_PROMPT,
     'question': QUESTION_PROMPT,
     'modes': PROMPT_MODES,
+    "question_marker": "Question:",
 }
 OPENAI = 'openai'
 ANTHROPIC = 'anthropic'
@@ -189,6 +195,10 @@ def mode_prompt(self):
     def question_prompt(self):
         return self.prompts['question'] + self.mode_prompt
 
+    @property
+    def question_marker(self):
+        return self.prompts['question_marker']
+
     @property
     def context_tokens(self):
         """The max number of tokens to be used for the context"""